Field Guide to Fable — Thariq Shihipar, Anthropic  🔗 **Source:** [Field Guide to Fable — Thariq Shihipar, Anthropic](https://www.youtube.com/watch?v=9fubhllmsBU#transcript) > [!note]+ Document Notes > tags: [ai-agents, claude-code, agentic-coding, prompt-engineering, anthropic-fable, human-ai-collaboration] > author: Thariq Shihipar > source: YouTube — AI Engineer (conference talk) > reference: https://www.youtube.com/watch?v=9fubhllmsBU > date created: 2026-07-06 12:44 > > # 🏷️ Title: Field Guide to Fable — Thariq Shihipar, Anthropic > <!-- deep-read:sections start --> ## 🧭 Table of Contents - [📝 Summary](#-summary) - [📖 Main Content](#-main-content) - [🗒️ Notes](#️-notes) - [📚 References](#-references) - [🗣️ Paraphrase](#️-paraphrase) - [📰 Newsletter](#-newsletter) - [✅ Action Items](#-action-items) - [🧪 Hypotheses](#-hypotheses) - [📊 Methodology](#-methodology) - [❓ Questions](#-questions) - [🔮 Future Research](#-future-research) - [💡 Implications](#-implications) - [🧠 Learning Styles Assessment](#-learning-styles-assessment) - [📊 Accessibility Assessment](#-accessibility-assessment) - [🌍 DEIB Assessment](#-deib-assessment) - [✅ Logical Validity](#-logical-validity) - [🛡️ DARVO Analysis](#-darvo-analysis) - [🎯 Conclusion](#-conclusion) ### 🔗 References - [Field Guide to Fable — Thariq Shihipar (YouTube / AI Engineer)](https://www.youtube.com/watch?v=9fubhllmsBU) - [On the Biology of a Large Language Model — Anthropic](https://transformer-circuits.pub/) — the interpretability paper he recommends - [Claude Code](https://www.anthropic.com/claude-code) — the harness he builds on - Fireside chat referenced: Thariq + Kat Woo + Simon Willison (same event, "12:30") ## 📝 Summary - **What happened.** Thariq Shihipar (member of technical staff at Anthropic, works on Claude Code) gives a conference "field guide" to **Fable**, a new Anthropic model class rolling out the same day. He structures the talk in four parts — *unhobbling Claude, finding your unknowns, dealing with the grief, and being unreasonable* — mixing concrete agentic-coding technique with a personal, almost philosophical reflection on what these models change. - **Why it matters.** It's a first-party, practitioner view of how to work with a frontier agentic model: not "better prompts" but a shift in mental model — treat the model as *grown, not designed*, discover its "capability overhang," and change your own workflow (find your unknowns, stay in the loop) to match a model that can now traverse far more of a problem autonomously. - **Key stakeholders.** AI engineers and agentic-coding practitioners (the audience), Anthropic (positioning Fable and Claude Code), and anyone building on or around these harnesses who needs to update their working assumptions as capability jumps. - **Risks and tradeoffs.** The advice is experiential and Anthropic-internal (a promotional context for Fable/Claude Code), so it's aspirational and hard to falsify. The "be unreasonable / tradeoffs aren't real" thesis is motivating but risks overreach; the honest counterweight he offers is "building is easier, but generating value is still hard." - **Bottom line.** Working with a Fable-class model is less about writing the perfect spec and more about (1) understanding and unleashing the model's spiky new capabilities, (2) using the model itself to surface your own unknown unknowns, (3) accepting the emotional loss of hand-craftsmanship, and (4) becoming more ambitious because many old tradeoffs no longer bind. "Go explore, make it real, and be less reasonable." ## 📖 Main Content - **Framing.** Fable is a "you'll remember where you were" model (he ranks the feeling with Sonnet 3.5-new, Opus 4, Opus 4.5). His metaphor: the RPG tutorial ends and the *open world* opens — exciting but intimidating because there's so much you can now do. The talk is a field guide to that open world. - **Part 1 — Unhobbling Claude.** Core premise: *models are grown, not designed.* Anthropic gives data, feedback, and compute; the result is organic, and **what contains the model is us** — the harness and prompts, a function of our understanding. "Unhobbling" = understanding Claude better to unleash it. **Capability overhang**: models get smarter in *spiky* ways. Example — asking which Pokémon end in "AW": a plain chat model fails, but Claude Code *fetches every Pokémon and writes a filter script* and finds Croconaw and Drednaw. Progression examples: chat needed pasted context → giving the model "arms" (bash tool) so it can build/search its own context (the insight behind Claude Code) → "Claude Tag," which can *proactively wake itself up* and work in multiplayer (the next wave of agents). Concrete shift: they **removed ~80% of Claude Code's system prompt**. System-prompt best practice evolved from *small prompt + few tools + many examples* (3.5-era) → *larger prompt + many examples + many tools* → now *smaller prompt, fewer examples* because examples **constrain** a model that's "more imaginative than the examples," and you should give **context, not constraints** (avoid "do not do this"). The **"ask-you-a-question" tool** shows the same curve: barely worked on Opus 4; by Opus 4.5 it could interview you with 40 questions; on Opus 4.8/Fable it can build a whole **HTML report with questions embedded**. Likewise Markdown→plan-mode→rich HTML reports. His meta-claim: this is **"closer to biology than physics"** — empirical and organic, with intuition to build (he recommends Anthropic's "biology of a large language model" paper). - **Part 2 — Finding your unknowns.** *The map is not the territory.* Your plan/prompt/spec is the map; the real codebase and constraints are the territory. When Claude hits something in the territory that's not in your map, that's an **unknown** — an unspecified decision point. Fable traverses so much ground that it hits *many* unknowns, so your bottleneck becomes matching map to territory. He uses a **knowns/unknowns matrix** (known-knowns = your prompt; known-unknowns; unknown-knowns = obvious-but-unwritten, know-it-when-you-see-it; unknown-unknowns). Then he uses the model to find them: - **Blind-spot pass** — "do a blind-spot pass to help me figure out my relevant unknown unknowns and prompt better"; have it read a module, search git diff/Slack for gotchas; also to learn new fields (he used it for color grading in video editing). - **Brainstorms & prototypes** — for unknown-knowns/design: "I have no visual taste — make an HTML page with four widely different design decisions so I can react to them." - **Interviews** — ask Claude to interview you; "prioritize questions that would change the architecture." - **References** — give it *another map*: existing code (any language) or an HTML mockup as the reference instead of writing a spec. - **Implementation notes** — have Fable log unknowns as it hits them so you can see and understand deviations. - **Quiz** — have Fable quiz you on what happened so you truly understand the work and can represent it at PR/merge time. The throughline: **stay in the loop.** - **Part 3 — Dealing with the grief.** First use of a Fable-class ("Mithril class") model brought both huge gain and a **sense of loss**. Coding before LLMs now feels like "a foreign country." He ran a ~30-person YC startup constantly forced into trade-offs because code was hard; revisiting that codebase, weeks of work now take hours. "How can you not laugh? Also, how can you not cry?" He loved hand-writing code (seeing and "rotating" the codebase in his mind) but also remembers "swimming in failure" — most projects fail, most startups go bankrupt. He can't go back. **"The only way out is through"** — stay in the loop, unhobble it, and come out with much more. - **Part 4 — Being unreasonable.** An Anthropic cultural belief: **"tradeoffs are not real."** Instead of reasonably prioritizing X against Y, ask "what if you just did all of it?" and **force reality to show you the tradeoff.** Good/fast/cheap → "now it's *pick three*." The way to do ambitious work is to make *yourself* more ambitious: "the only way to prove that agents work is to do the best work of our lives faster than ever before." (He built the talk's deck in ~4 hours with Fable.) His resolution: be more productive but **work less** and spend time with people he cares about. Honest caveat: **"building is easier, but generating value is still hard"** — the point is value, and finding it takes many swings. Closing charge: "go explore, make it real, and be less reasonable." - **Contested / uncertain.** Much is first-party and unfalsifiable (an Anthropic engineer promoting Anthropic models on launch day). "Tradeoffs are not real" is a motivational overstatement he himself qualifies. The techniques are concrete and testable; the capability and cultural claims are self-reported. ### 🗒️ Notes #### ⏱️ [00:02:07](https://www.youtube.com/watch?v=9fubhllmsBU&t=127) The four parts - Fable rolling out "later today"; a "you'll remember it" model (ranks with Sonnet 3.5-new, Opus 4, Opus 4.5). - Metaphor: RPG tutorial → open world. Four parts: unhobbling Claude, finding your unknowns, dealing with the grief, being unreasonable. #### ⏱️ [00:02:42](https://www.youtube.com/watch?v=9fubhllmsBU&t=162) Unhobbling — grown, not designed - Models are grown, not designed; what contains them is us (harness + prompts). - Unhobbling = understand Claude better to unleash it. #### ⏱️ [00:04:27](https://www.youtube.com/watch?v=9fubhllmsBU&t=267) Capability overhang - Spiky intelligence. Pokémon-ending-in-"AW": chat model fails; Claude Code writes a filter script and finds Croconaw + Drednaw. - Progression: chat needs pasted context → "arms" (bash tool) to build its own context (→ Claude Code) → Claude Tag wakes itself up, works multiplayer. #### ⏱️ [00:06:00](https://www.youtube.com/watch?v=9fubhllmsBU&t=360) Smaller prompts, fewer examples - Removed ~80% of Claude Code's system prompt. - Best practice evolved: small+examples (3.5) → large+many tools → now smaller, fewer examples (examples constrain), give context not constraints, avoid "do not." #### ⏱️ [00:07:05](https://www.youtube.com/watch?v=9fubhllmsBU&t=425) Ask-a-question tool + rich output - Opus 4 barely called it → Opus 4.5 could interview with 40 questions → Opus 4.8/Fable builds HTML report with embedded questions. - Output curve: Markdown → plan mode (for you) → in-depth HTML reports. #### ⏱️ [00:08:30](https://www.youtube.com/watch?v=9fubhllmsBU&t=510) Biology, not physics - Treat the model as empirical/organic; build intuition. Recommends "biology of a large language model." #### ⏱️ [00:09:18](https://www.youtube.com/watch?v=9fubhllmsBU&t=558) Finding unknowns — map ≠ territory - Map = plan/prompt/spec; territory = real codebase. Anything in the territory not in the map = an unknown (an unspecified decision). - Fable covers so much ground it hits many; your bottleneck is matching map to territory. Matrix: known-knowns / known-unknowns / unknown-knowns / unknown-unknowns. #### ⏱️ [00:11:00](https://www.youtube.com/watch?v=9fubhllmsBU&t=660) Techniques to surface unknowns - Blind-spot pass (find unknown unknowns; learn new fields — e.g. color grading). - Brainstorms/prototypes (4 divergent HTML designs to react to). - Interviews ("prioritize questions that would change the architecture"). - References (give it another map: code/HTML mockup instead of a written spec). - Implementation notes (log unknowns as it hits them). - Quiz (have Fable quiz you so you can represent the work at PR time). Throughline: stay in the loop. #### ⏱️ [00:14:40](https://www.youtube.com/watch?v=9fubhllmsBU&t=880) The grief - Gain + loss. Pre-LLM coding feels like "a foreign country." Weeks → hours. "How can you not laugh? Also, how can you not cry?" - Loved hand-coding but remembers "swimming in failure"; can't go back. "The only way out is through." #### ⏱️ [00:16:43](https://www.youtube.com/watch?v=9fubhllmsBU&t=1003) Being unreasonable - "Tradeoffs are not real." Force reality to show the tradeoff. Good/fast/cheap → pick three. - Prove agents work by doing "the best work of our lives faster than ever before." Built the deck in ~4 hours. - Caveat: "building is easier, but generating value is still hard." Resolution: more productive, work less, more time with people. "Go explore, make it real, be less reasonable." ## 📚 References - [Field Guide to Fable — Thariq Shihipar, AI Engineer (2026-07-06)](https://www.youtube.com/watch?v=9fubhllmsBU) — primary source - [On the Biology of a Large Language Model (Anthropic / Transformer Circuits)](https://transformer-circuits.pub/) — the interpretability paper he recommends - [Claude Code](https://www.anthropic.com/claude-code) — the harness under discussion - [Anthropic — Claude models overview](https://www.anthropic.com/claude) — for the Sonnet/Opus/Fable lineage he references ## 🗣️ Paraphrase - **Open.** Fable is a landmark model; the RPG tutorial is over and the open world begins — thrilling and intimidating. - **Unhobble.** Models are grown, not designed; we constrain them via harness and prompts. Unhobbling means understanding Claude to release its spiky "capability overhang" (Pokémon script example). Newer models want smaller prompts, fewer examples, and context over constraints; treat it like biology, not physics. - **Unknowns.** The map (your spec) isn't the territory (the real code). Gaps are "unknowns"; Fable hits many, so use it to find them — blind-spot passes, divergent prototypes, letting it interview you, giving it reference maps, logging implementation notes, and having it quiz you. Stay in the loop. - **Grief.** These models bring huge gain and real loss; hand-coding is receding into "a foreign country," but there's no going back — the only way out is through. - **Unreasonable.** At Anthropic, "tradeoffs are not real"; be less reasonable, attempt all of it, and force reality to reveal the real constraints. Building is easier now, but generating value is still hard — so aim it at value, work ambitiously, and (his resolution) work less while doing more. ## 📰 Newsletter **A field guide to working with frontier agentic models — from the person who builds the harness.** Anthropic's Thariq Shihipar (Claude Code) gave a talk on "Fable" that's really a talk about *how your job changes* when the model can traverse a whole problem on its own. Four ideas worth stealing: (1) **models are grown, not designed** — "what contains them is us," so your prompts and harness are the real lever (they just deleted ~80% of Claude Code's system prompt because newer models want *context, not constraints*); (2) **the map is not the territory** — the gaps between your spec and the real codebase are "unknowns," and you can have the model surface your own blind spots (blind-spot passes, letting it interview you, giving it a reference to copy instead of a spec); (3) **the grief is real** — hand-craftsmanship is receding and "the only way out is through"; and (4) **be unreasonable** — "tradeoffs are not real," so aim higher, but remember "building is easier, generating value is still hard." For anyone doing loop-engineering or agentic work, this is the philosophy behind the practice — 19 minutes, and the "stay in the loop" throughline is the whole maker/checker discipline in a sentence. 🧭🤖 ## ✅ Action Items #### DevRel - Developer advocacy: adapt the "map is not the territory / find your unknowns" framing into a talk or workshop on agentic coding. - Education: teach the six unknown-surfacing techniques (blind-spot pass, prototypes, interviews, references, implementation notes, quiz) as a concrete curriculum. - Developer experience: publish a "context, not constraints" prompt/skill guide; audit existing SKILL.md files for over-constraining examples. - Developer marketing: use "grown, not designed" and "capability overhang" as framings for how tools improve. - Community management: run a "find your unknowns" community exercise. - Internal/external/cross-project: standardize the "have the model quiz you before PR" habit for staying in the loop. #### OSPO - Policy & Governance: as models traverse more autonomously, define oversight/in-the-loop requirements for agent-run work. - License Compliance: n/a to this source. - Security Management: "capability overhang" cuts both ways — agents with tools can do more, so scope tool access deliberately. - Software Inventory: n/a. - Developer Support: office hours on agentic-coding workflow (unknowns, references, implementation notes). - Community Engagement: share the "biology not physics" empirical mindset for evaluating models. - Strategic Alignment: treat "generating value is still hard" as a reminder to measure outcomes, not build velocity. #### Community - Maintainers: adopt reference-driven prompting (give the agent an existing implementation as the map). - Contributors: use blind-spot passes before large PRs. - Sponsors: fund tooling that keeps humans in the loop (implementation notes, quizzes). - Ecosystem: shared library of "find your unknowns" prompts/skills. #### WordPress - Core: n/a directly; relevant to how AI-assisted core contributions should keep humans in the loop. - Core Performance: n/a. - Core AI: this talk is on-charter — the "context not constraints" and "capability overhang" lessons should inform WordPress AI tooling and evaluation. - **Core Program Team:** adopt "have the model interview/quiz you" as a spec-hardening ritual for AI-assisted planning. - Mobile / Playground: build "four divergent designs to react to" prototyping into design workflows. - Community / Training: Learn WordPress module on agentic-coding technique (unknowns, references). - Marketing: "grown not designed" as an honest way to explain model behavior to non-technical audiences. - Support / Design: use divergent-prototype technique for design decisions; keep a human reviewer in the loop. - Documentation: prefer reference examples over long constraint-heavy prompts in AI docs. - Accessibility: when the model builds HTML reports/prototypes, hold them to WCAG (don't let speed skip a11y). - Meta / CLI / Hosting / Tide / Openverse / Themes / Plugins: n/a directly to this source. #### The WPCC - Programs: teach agentic-coding literacy (the four-part framework) to maintainers. - Projects: a "stay in the loop" governance pattern for AI-assisted project work. - Funding experiments: n/a. ### FAIR - Technical independence: "what contains them is us" — control of harness and prompts is a form of independence worth preserving locally. - Policies and covenants: human-in-the-loop covenants for autonomous agent work. - Productized governance artifacts: a "find your unknowns" checklist as a reusable spec-hardening artifact. ## 🧪 Hypotheses - **H1.** For Fable-class models, smaller system prompts with fewer examples produce better/more-imaginative outputs than large example-heavy prompts (reversing earlier best practice). - **H2.** Using the model to surface unknowns (blind-spot pass, interviews) before implementation reduces downstream rework versus prompting straight into implementation. - **H3.** Giving the model a *reference implementation* as the "map" yields higher first-pass correctness than an equivalently detailed written spec. - **H4.** As model capability rises, the human bottleneck shifts from writing code to specifying intent and adjudicating unknowns (map–territory matching). - **H5.** Practitioners who "stay in the loop" (implementation notes + quiz-back) produce fewer defects and better-understood PRs than those who accept large autonomous diffs unreviewed. - **H6.** Framing effects ("be unreasonable / attempt all of it") measurably increase the ambition/scope of what practitioners attempt with capable models — with a corresponding rise in both value and wasted effort ("many swings"). ### 📊 Methodology - **The talk's basis (informal):** first-party experiential report from an Anthropic engineer who builds Claude Code, illustrated with anecdotes (Pokémon capability example, the ask-a-question tool's progression across model versions, a deck built in ~4 hours) and internal practice. Not an empirical study. - **To test H1:** A/B prompt experiments (small-vs-large system prompts) on identical agentic tasks, scored on task success and solution quality. - **To test H2/H3:** controlled coding tasks comparing unknown-surfacing vs. direct prompting, and reference-map vs. written-spec, measuring rework and first-pass correctness. - **To test H4:** time-motion study of practitioner activity (writing code vs. specifying/adjudicating) across model generations. - **To test H5:** defect-rate and comprehension measures for in-the-loop vs. hands-off review styles. - **To test H6:** framing-manipulation study on attempted scope and realized value. - **Confounders:** launch-day promotional context (incentive to present Fable favorably), single-practitioner anecdotes, rapid tool changes, and selection bias in a conference audience. ### ❓ Questions - Where exactly does "context, not constraints" break down — do safety-critical or compliance contexts still need explicit "do not" rules? - How do you keep a human genuinely "in the loop" when the model can produce a whole HTML report or large diff faster than you can review it? - Is "capability overhang" a stable property to design around, or a moving target that makes today's techniques obsolete quickly? - When is "be unreasonable" productive ambition vs. scope-creep that burns swings without generating value? - What's the failure mode of over-relying on the model to find your unknowns (does it surface *its* blind spots, not yours)? ### 🔮 Future Research - Empirical validation of the smaller-prompt/fewer-examples shift across model generations and domains. - Best practices and tooling for human-in-the-loop review when output outpaces reading speed. - Measurement of the map–territory "unknowns" workflow: does it demonstrably reduce rework? - Study of framing/ambition effects on realized value vs. wasted effort with capable agents. - Interpretability-informed "field guides": can the "biology" research directly inform prompting technique? ### 💡 Implications - **Practice.** Update prompting toward context over constraints; use the model to surface your unknowns; give references instead of specs; stay in the loop with implementation notes and quiz-backs. - **Policy.** Autonomous traversal raises the stakes on human oversight — build in-the-loop checkpoints as capability grows. - **Security.** Tool access is where "capability overhang" is realized; scope tools intentionally. - **Education.** Agentic-coding literacy is now a distinct skill set (unknowns, references, staying in the loop) worth teaching explicitly. - **Economics/wellbeing.** "Building is easier, value is still hard" reorients effort toward outcomes; the "work less, do more, be with people" resolution is a humane counterpoint to hustle framing. ### 🧠 Learning Styles Assessment - Reading level estimate: approximately Flesch–Kincaid Grade 8–10 spoken register (conversational talk with metaphors and some jargon). Flesch Reading Ease roughly 55–65. #### 👀 Visual learners - Present: implied slides (the deck he built) with examples and likely diagrams (matrix, progressions) — visible in the video, not the transcript. - Improve: publish the knowns/unknowns matrix and the model-progression timeline as standalone diagrams. #### 🔊 Aural learners - Present: strong — a conference talk; the argument works fully as audio. - Improve: chapter markers per part; a tightened edit of filler. #### 🗣️ Verbal learners - Present: strong — memorable analogies ("open world," "map is not the territory," "biology not physics," "grief," "be unreasonable"). - Improve: a glossary (capability overhang, unhobbling, unknown unknowns, harness). #### 🧩 Kinesthetic learners - Present: strong — concrete, copy-able techniques (blind-spot pass, interviews, references, implementation notes, quiz). - Improve: a hands-on worksheet to run each technique on a real task. #### 🧠 Logical learners - Present: strong — a clean four-part structure and a 2×2 knowns/unknowns matrix. - Improve: show data behind capability claims; a decision guide for which unknown-surfacing technique to use when. #### 🧑‍🤝‍🧑 Social learners - Present: partial — a live talk with an opening group selfie; references a fireside chat. - Improve: discussion prompts and a community challenge to try the techniques. #### 🧘‍♀️ Solitary learners - Present: strong — reflective, personal ("the grief") and self-paced techniques. - Improve: journaling prompts on your own gain/loss and your unknowns. #### 🌿 Naturalistic learners - Present: strong — the "grown, not designed / biology not physics / organic" framing is an explicit natural-systems metaphor. - Improve: extend the biology metaphor into an evaluation practice (observe, hypothesize, test). ### 📊 Accessibility Assessment #### Screen reader semantics - Present: YouTube auto-captions (used here) and standard player semantics. - Improve: publish a human-corrected transcript — auto-captions mangle names ("Tarik Shaupar" for Thariq Shihipar) and terms ("on hobble" for "unhobble," "Sable"/"Mithril" for Fable). #### Keyboard navigation - Present: keyboard-operable player. - Improve: n/a beyond platform. #### Contrast and readability - Present: depends on slide design. - Improve: ensure slide text (code, matrices) meets contrast and is legible on mobile. #### Text resizing and reflow - Present: platform-controlled. - Improve: provide a text write-up (he says it's a planned blog series) for reflow/zoom. #### Media and captions - Present: auto-captions only. - Improve: corrected closed captions + downloadable transcript; the jargon/name errors are a real gap. #### Forms and interactions - Present: n/a (video). - Improve: n/a. #### Standards reference - Gaps map to WCAG 2.2 AA: 1.2.2 Captions (auto-caption accuracy), 1.2.1/1.2.3 (text alternative/transcript), 1.4.3 Contrast (slides). ### 🌍 DEIB Assessment #### Cultural breadth - Present: limited — US tech-conference context, English-only, Anthropic-internal vantage. - Improve: acknowledge access barriers (frontier-model cost) and non-Western engineering contexts. #### Inclusive language - Present: good — warm, first-person, non-gatekeeping; jargon mostly explained via analogy. - Improve: a glossary would widen access beyond the AI-engineer in-group. #### Neurodiversity support - Present: strong — clear four-part scaffold, repeated signposting, concrete examples. - Improve: an up-front summary/agenda slide and consistent per-part recap. #### Representation - Present: single-speaker, single-company perspective; names a few colleagues (Kat Woo, Simon Willison). - Improve: pair with independent/critical perspectives on frontier-model claims. #### Psychological safety - Present: notably strong — emotionally honest ("the grief," "swimming in failure," fear of loss), which models vulnerability and normalizes ambivalence about AI. - Improve: balance the "be unreasonable / best work of our lives" ambition with explicit anti-burnout framing (he partly does with "work less, be with people"). ### ✅ Logical Validity #### Strong reasoning - **Analogical, strong and generative.** "Map is not the territory," "open world," "biology not physics," and "grown, not designed" are apt analogies that each do real explanatory work (locating the human bottleneck, the empirical stance, the harness-as-constraint). - **Inductive from concrete progression, reasonable.** The ask-a-question tool across Opus 4 → 4.5 → 4.8/Fable is a specific, checkable capability curve supporting the "spiky improvement" claim. - **Constructive/actionable.** The unknowns matrix plus six named techniques translate the philosophy into testable practice. #### Fallacies - **Anecdotal evidence.** Location: capability claims rest on selected anecdotes (Pokémon script, 4-hour deck, personal codebase revisit). Weakness: unrepresentative and unmeasured. Fix: cite benchmarks or controlled comparisons. - **Conflict of interest / promotion.** Location: an Anthropic engineer extolling a same-day-launch Anthropic model and Claude Code. Weakness: incentive to present favorably; "you'll remember this model" is marketing framing. Fix: independent evaluation; he is at least transparent about affiliation. - **Overgeneralization / hyperbole.** Location: "tradeoffs are not real" and good/fast/cheap "pick three." Weakness: literally false as stated; tradeoffs are relaxed, not abolished. Mitigation: he self-corrects with "building is easier, but generating value is still hard." - **Survivorship framing (mild).** Location: "weeks → hours" retrospective on his old codebase. Weakness: highlights wins; the tasks that *don't* compress this way are absent. Fix: acknowledge where the speedups don't hold. #### Tone of voice - Warm, reflective, and evangelistic but self-aware — equal parts practitioner tutorial and personal essay. Confident and inspirational about capability, genuinely vulnerable about loss, and lightly hedged about value ("generating value is still hard"). Community-oriented ("go explore… be less reasonable"). ### 🛡️ DARVO Analysis - **Deny:** Not applicable — no accusation is being answered; this is a talk, not a defense. - **Attack:** Not applicable — no attack on critics or opponents. - **Reverse victim/offender:** Not applicable — no victim/offender dynamic. The nearest emotional beat ("the grief," "swimming in failure") is self-directed reflection, not blame-shifting. - Overall: no DARVO pattern. The relevant caution is straightforward promotional conflict of interest, not manipulation. ## 🎯 Conclusion - **Bottom line.** Fable (and Fable-class models) shift the work from crafting perfect specs to *unleashing spiky new capability, surfacing your own unknowns, and staying in the loop while the model does far more on its own.* The mindset is empirical ("biology, not physics"), the technique is concrete (six ways to find unknowns), the emotion is real ("the grief"), and the ambition is deliberately raised ("be unreasonable") — tempered by the honest reminder that building is now easy but generating value is not. - **Practical next steps for readers.** (1) Rewrite prompts/skills toward context over constraints; cut over-constraining examples. (2) Before implementing, run a blind-spot pass and let the model interview you; give it a reference implementation instead of a spec. (3) Have it log implementation notes and quiz you before you merge — that's staying in the loop. (4) Aim higher than feels reasonable, then point that ambition at *value*, not just building. (5) Watch for the honest limits — human review capacity, promotional framing, and the tasks that don't compress. ## 🔄 Refactor Appendix — Field Guide to Fable > Density: 2–5 Literature atoms (H3), 0–3 Permanent candidates (H4). > Defaults: 3 Literature, 1 Permanent. Resonance filter: if you can't > articulate in one line why a claim matters, drop it. ### 📚 Literature Atoms (H3) #### [H1] Models are grown, not designed; the harness contains them > [!quote] ID:rw_fable_grown1 > "the models are grown, not designed… what contains them is us, right? The harness we put them in and the way we prompt them" **Context & use**: - Timestamp: ~02:42 (Part 1, unhobbling). - Why this matters: reframes the human's leverage — improving prompts/harness ("unhobbling"), not the weights, is how most practitioners unlock more capability. **Tags**: #xref #aphorisms **Highlight**: [view highlight](https://www.youtube.com/watch?v=9fubhllmsBU&t=162) <!-- atomicity: 1 idea ✓ --> #### [H2] Newer models want context, not constraints > [!quote] ID:rw_fable_context2 > "this new class of models want… a smaller system prompt. The examples tend to constrain it… we tried to give it context and not just constraints." **Context & use**: - Timestamp: ~06:00 (removed ~80% of Claude Code's system prompt). - Why this matters: reverses prior prompt best-practice — fewer examples, less "do not," more context. Directly actionable for SKILL.md and prompt design. **Tags**: #xref #addendum **Highlight**: [view highlight](https://www.youtube.com/watch?v=9fubhllmsBU&t=360) <!-- atomicity: 1 idea ✓ --> #### [H3] The map is not the territory; gaps are "unknowns" > [!quote] ID:rw_fable_map3 > "the plan and prompt and spec that I have in my mind is the map… the territory is the actual code base… whenever Claude runs into something in the territory that's not in the map, I call that an unknown" **Context & use**: - Timestamp: ~09:18 (Part 2, finding your unknowns). - Why this matters: names the new human bottleneck with capable agents — matching spec to reality — and motivates using the model to surface unknown unknowns. **Tags**: #xref **Highlight**: [view highlight](https://www.youtube.com/watch?v=9fubhllmsBU&t=558) <!-- atomicity: 1 idea ✓ --> #### [H4] Building is easier, but generating value is still hard > [!quote] ID:rw_fable_value4 > "building is easier, but generating value is still hard" **Context & use**: - Timestamp: ~18:28 (Part 4, being unreasonable). - Why this matters: the honest counterweight to "tradeoffs aren't real" — reorients effort from build-velocity to outcomes; the crucial caveat on all agentic-productivity hype. **Tags**: #xref #aphorisms **Highlight**: [view highlight](https://www.youtube.com/watch?v=9fubhllmsBU&t=1108) <!-- atomicity: 1 idea ✓ --> ### 🧠 Permanent Candidates (H4) #### [P1] With capable agents, the human's job is to specify intent and adjudicate unknowns **Claim** (one sentence, Courtney's voice): As a coding agent becomes able to traverse most of a problem itself, the human's leverage moves from writing the code to specifying intent well and resolving the unspecified decision points ("unknowns") the agent surfaces. **Evidence pointer**: [[#H3]] **Scope & conditions**: Holds for capable agentic-coding harnesses on real codebases. Breaks when the task is small enough that specification overhead exceeds doing it directly, or when the agent's autonomy outpaces the human's ability to review. **Links**: - [[Loop engineering OS]] — because applies-in maker/checker discipline - [[Map is not the territory]] — because extends - [[Stay in the loop]] — because evidence-for - [[Context not constraints]] — because extends <!-- atomicity: 1 idea ✓ --> ### ✏️ Fleeting Queue - [F1] Audit my SKILL.md files for over-constraining examples / "do not" rules — shift toward context per Fable guidance — skill library #fleeting #inbox - [F2] Try the six unknown-surfacing techniques (blind-spot pass, interview-me, reference-as-map, implementation notes, quiz-me) on a real courtneyr.dev / Outpost task #fleeting #inbox - [F3] "Stay in the loop" = the maker/checker throughline; cross-link to loop-engineering OS and the agentic-WordPress governance gap #fleeting #inbox ### → Things 3 (max 2 actions) - [ ] Run a blind-spot pass + "interview me" on the next substantial WordPress spec before implementing #things - [ ] Review skill library prompts for "context not constraints" — trim over-specified examples #things <!-- deep-read:sections end --> 📄 **Full text:** [[Full Document Contents/Articles/Field Guide to Fable — Thariq Shihipar, Anthropic.md|Field Guide to Fable — Thariq Shihipar, Anthropic by Thariq Shihipar]] ## Highlights - So, first, unhobbling Claude. I think something we say really often is that the models are grown, not designed, right? We don't wake up and be like, "We need 99% on SWE-bench," right? Like, the models are, you know, something we we grow carefully. We give it data and feedback and compute. Um but, ultimately, it's, you know, something that we it's a little bit organic, and we sort of figure out and learn with the model as we use it. ([View Highlight](https://read.readwise.io/read/01kww5etdbvgf97nyc0nbxt01z)) - Note: Core reframe: models are grown, not designed. What contains them is the harness + prompts — so 'unhobbling' (better prompts/harness) is the human's real leverage. - Tags: aphorisms, atom/literature/h1 - And then as the models get smarter, you can give it more information and more instructions, and they start following them. And so, it's a larger system prompt with lots of examples and many tools, right? But most recently, we found this new class of models want fewer what want a smaller system prompt. The examples tend to constrain it cuz it's actually more imaginative than the examples we give it. ([View Highlight](https://read.readwise.io/read/01kww5ewqt3t2hpzr69bqv9hnr)) - Note: Prompt best-practice reversal: newer models want smaller prompts + fewer examples (examples constrain) — give context, not constraints. Directly actionable for SKILL.md design. - Tags: atom/literature/h2 - And so one of the things that I think a lot about is that the map is not the territory, right? When I'm working on a coding problem, the plan and prompt and spec that I have in my mind is the map, right? But the territory is the actual code base, the real world, the constraints that Claude needs to navigate, right? ([View Highlight](https://read.readwise.io/read/01kww5ez4zm5fw54jpf1j9napy)) - Note: The new human bottleneck with capable agents: matching your spec (map) to reality (territory). Gaps = 'unknowns' the model can help you surface. - Tags: atom/literature/h3 - Uh I think it's also worth calling out that building is easier, but generating value is still hard. And I think this is something that we run into, you know, as AI engineers sometimes where we think so much about the process of building and our our setups. Um but the the point is to generate value, right? ([View Highlight](https://read.readwise.io/read/01kww5f1r7m37p7g100c8qs6s5)) - Note: The honest counterweight to 'tradeoffs aren't real': reorients effort from build-velocity to outcomes. The crucial caveat on agentic-productivity hype. - Tags: aphorisms, atom/literature/h3